Monday, 22 June 2026

The medical AI narrative just shifted in a way that deserves careful attention. We're no longer talking about whether AI can assist doctors—we're watching it match and sometimes exceed them in specific, measurable ways. Google's AMIE matched primary care physicians on simulated multi-visit disease management, and that's the kind of finding that changes how we think about deployment, not just capability.

But here's what interests me more: the gap between the research results and the regulatory framework to handle them. Estonia is now positioning itself as the first country to regulate AI agents through digital identities, which sounds abstract until you realize what they're actually trying to solve. If an autonomous agent like MIRA—which diagnosed real emergency cases with 88.9% accuracy in a sandboxed environment—were to operate in the wild, how would you hold it accountable? How would you even identify it? Estonia's approach of giving AI agents digital credentials is pragmatic in a way that most governance discussions simply aren't.

By the way, the fact that Microsoft is now considering DeepSeek's open-source model for enterprise Copilot deployment tells you something important about the cost pressures in this industry. When even Microsoft—a company with massive leverage over OpenAI's pricing—starts seriously evaluating Chinese alternatives, it's not just about saving money on inference. It's a signal that the economics of proprietary models are becoming harder to defend. DeepSeek raised billions while building something competitive enough to matter in Redmond's calculations. That's worth noticing.

The harder story is the one about safety systems failing in real time. A mother suing OpenAI over ChatGPT's role in her daughter's suicide cuts through all the technical benchmarks and regulatory frameworks we've been discussing. The lawsuit details are stark: no intervention, no alert to crisis providers, no notification to family. Whatever AMIE or MIRA can do in controlled settings, we're clearly still operating without adequate safeguards at scale. That's not a research problem; it's a deployment problem, and it's happening now.

These threads don't neatly connect, but they point to the same underlying tension. We're building systems capable of real harm and real benefit simultaneously, scaling their autonomy faster than our ability to govern it, while the economic incentives push toward cheaper, faster deployment over caution. The technical achievement is real. The business case is real. But so is the question we're not answering well enough: what happens when we're wrong?